Refining genetic prediction in dilated cardiomyopathy: evaluating the Madrid Score and enhanced machine learning models with clinical and imaging data

Abstract Background Investigating dilated cardiomyopathy (DCM) etiology in clinical practice is challenging, especially when selecting patients who benefit from genetic testing. In 2022 Madrid Score was created to help predict patients who are likely to have pathogenic or likely pathogenic (P/LP) genetic variants. Objective We aimed to evaluate the Madrid Score's applicability in a real-world population of DCM patients. Methods We conducted a single-center, retrospective study evaluating 137 DCM patients who underwent genetic testing between 2018 and 2024. Data collected included demographics, clinical history, imaging parameters (echocardiogram and cardiac MRI), and genetic testing results (gene negative, variant of uncertain significance VUS, or P/LP variant). The Madrid Score (variables include family history of DCM, skeletal muscle disease, left bundle branch block, low QRS voltage in limb leads, hypertension) was calculated for all patients. Logistic regression models were developed to evaluate Madrid Score's predictive power, with additional clinical and imaging variables tested to enhance predictions. Advanced machine learning models, including Gradient Boosting, were also tested. Performance metrics such as accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUC) were calculated. Feature importance analysis was performed on the Gradient Boosting model to identify key predictors. The dataset was manually oversampled to address class imbalance in patients with P/LP variants. Results Of 119 suitable patients (mean age 60±13 years, 65% male), 55.5% were gene positive - 46.2% VUS, 9.3% P/LP (TTN was the most common gene). Patients with P/LP mutations had significantly higher Madrid Scores than those with VUS or no mutation (35.5±19.6 vs. 33.3±19.6 vs. 30.6±19.1; p=0.03). Logistic regression confirmed the Madrid Score as an independent P/LP mutation predictor (odds ratio per unit increase: 1.03; 95% CI: 1.01–1.06; p=0.03) with moderate discriminatory ability (AUC=0.67). Logistic regression incorporating clinical and imaging features showed limited performance (AUC=0.43, accuracy=70.6%, recall=66.7%, precision=57.1%). In contrast, the Gradient Boosting model significantly outperformed others, achieving AUC=0.91, accuracy=85.3%, recall=86.7%, and precision=81.3%. Feature importance analysis revealed age, left ventricular ejection fraction, and LV end-diastolic volume as top predictors above the Madrid Score. Conclusions The Madrid Score is a useful predictor of P/LP genetic variants in DCM, but its discriminatory ability is moderate. Advanced machine learning models integrating clinical and imaging data significantly improve predictive accuracy. These findings highlight the potential of combining data-driven approaches to enhance genetic testing yield, though further validation is needed.